Near-Real-Time Integration of Multisource Precipitation Products Using a Multiscale Convolutional Neural Network

Merging multisource precipitation data based on deep learning models to create an accurate rainfall dataset has received significant interest in recent years. This article proposes a deep learning model to produce a high-accuracy, near-real-time precipitation product for the north-central region of Vietnam during the period 2019–23, with a spatial resolution of 0.048 and a temporal resolution of 1 h. The input multisource data including near-real-time satellite-derived precipitation products [PERSIANN-CCS, Global Satellite Mapping of Precipitation Near–Real Time (GSMaP-NRT), and IMERG-Early Run], radar precipitation, and gauge observations and spatial features neighbor estimation (NE) and precipitation occurrence probability (POP) are merged by a multiscale convolutional neural network (CNN)–based model with focal loss function and mean-square-error loss function for classification and regression tasks, respectively. Extensive experiments demonstrate that the proposed precipitation product outperforms all the input precipitation products and the post-real-time global precipitation products including GSMaP-moving vector with Kalman filter (MVK)-Gauge and IMERG-Final Run. It achieves classification metrics with a critical success index (CSI) of 0.65 and a BIAS of 1.03, with improvements from 31.58% to 54.8% in CSI and from 17.47%to 105.82% in BIAS, compared to radar, GSMaP-MVK-Gauge, and IMERG-Final Run products. For regression metrics, it achieves an RMSE of 3.34 mm h21, and a modified Kling–Gupta efficiency (mKGE) of 0.70, with improvements from 10.18% to 100% in RMSE and from 15.71% to 71.43% in mKGE over the same reference products. These results indicate that the merged product has a greater capability to detect rainfall events and significantly better overall performance, with lower systematic and random errors compared to the same reference products. Moreover, the proposed method outperforms the other methods, including random forest, long short-term memory, and the original multiscale CNN.